Google DeepMind · 2024 · 12 · 04 · Impact · ~2 min read
DeepMind's GenCast beat the world's best weather model
What's actually new
- Beat the world's best on 97% of tests. GenCast's 15-day ensemble forecasts more accurate than ECMWF's ENS model on 1,320 of 1,360 evaluation targets — Nature, December 2024.
- 8 minutes vs hours. A single global ensemble forecast takes ~8 minutes on a Google TPU. Traditional weather supercomputers take much longer for the same output.
- Better at extremes. Cyclones, heatwaves, renewable-energy yield forecasts — GenCast got noticeably better at the rare events that matter most.
- Open weights. DeepMind released GenCast publicly. Universities and weather services can run their own.
If you want more
Worth knowing
- Still depends on initial conditions from physics-based models. AI doesn't replace the satellites and buoys that feed it data — it just processes them more cleverly.
- Local, fine-scale forecasts (single-village rainfall) are still where physics-based models hold their ground.
- 'Beat ECMWF' headlines need translation. ECMWF themselves now run their own AI model (AIFS) alongside the traditional one — these aren't enemies, they're cooperators.
Who should care
Anyone whose work depends on weather forecasts — farmers, energy traders, shipping companies, emergency planners, sports event organisers. Climate scientists. People in regions hit by tropical cyclones or extreme heat. Anyone wondering 'what is AI actually good for in the real world?' — this is one of the cleanest answers.
What to do about it
Watch your local forecasts get better through 2025-2026 as GenCast and similar AI models feed into your weather app's pipeline (often invisibly). For a meaningful comparison, try ECMWF.int's open AI charts vs your favourite app — the AI ones are usually crisper.
Honest take
GenCast was the moment 'AI replaces a 50-year-old supercomputer pipeline' stopped being a thought experiment and became a Nature paper with hard numbers. The deeper story is what AI keeps doing in science: not replacing physics, but learning patterns from physics's data well enough to forecast faster and sometimes more accurately. Weather is the test case; protein folding, materials, fusion physics are next. The boring news for the public is that your weather app will quietly get better — and the credit isn't the app maker's, it's an AI model running deep underneath.
Sources
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com